NVIDIA's Tightrope Walk: Chips, Smugglers, and the China Paradox
In 2022, NVIDIA held 95% of China's AI chip market. By mid-2025, CEO Jensen Huang stated bluntly: "We are 100 percent out of China. We went from a 95% market share to 0%." Yet Chinese companies have placed orders for over 2 million NVIDIA H200 chips for 2026. ByteDance alone plans to spend 100 billion yuan on NVIDIA hardware this year.
How does a company go from total dominance to zero market share while still receiving billion-dollar orders? The answer lies in a policy landscape that defies simple logic.
The Regulatory Whiplash
The story begins in October 2022, when the Biden administration rolled out export controls cutting off China's access to high-end AI chips. The intent was to prevent the Chinese military from accessing cutting-edge processing power.
NVIDIA adapted by engineering the H20 chip specifically to fall within permitted limits. Chinese companies, meanwhile, optimized their systems around available hardware and accelerated domestic chip development.
Then came April 2025: the Trump administration banned even these compliant chips. Three months later, it reversed course, allowing H200 sales with a novel twist. The U.S. government would collect a 25% fee on those very sales, paid directly to the Treasury.
The policy created an unusual situation. On December 8, 2025, federal prosecutors announced Operation Gatekeeper, shutting down a $160 million smuggling network that had shipped banned NVIDIA chips to China. The next day, President Trump announced those same chips could now be legally exported with the 25% surcharge.
Defense attorneys immediately seized on the contradiction, arguing in court filings that the president's announcement undermined the prosecution's national security rationale.
The Smuggling Economy
The demand for restricted chips created a thriving gray market. The Center for a New American Security estimates between 10,000 and several hundred thousand AI chips were smuggled to China in 2024 alone. A black market emerged with premiums reaching 50% above list price.
Operation Gatekeeper revealed the mechanics: operatives entering the U.S. under false pretenses, front companies, warehouse operations relabeling NVIDIA GPUs with fake branding like "Sandkayan," export paperwork misclassifying the goods as generic "adapters."
Singapore emerged as a key transit point. The city-state's share of NVIDIA revenue jumped from 9% to 22% in two years. Investigations traced chips through Singapore to Malaysia and onward. Nine arrests followed raids on 22 locations. One investigation examined whether DeepSeek, the Chinese AI company whose model surprised the industry, obtained restricted chips through Singaporean intermediaries.
NVIDIA's largest Southeast Asian customer, Megaspeed International, became a focal point. The company purchased nearly $2 billion in NVIDIA products through its Malaysian subsidiary, but the scale of purchases raised questions about whether the hardware actually stayed in declared locations.
China's Response: Self-Sufficiency
China's reaction evolved from frustration to strategic pivot. In September 2025, Beijing ordered domestic tech giants to stop purchasing NVIDIA chips altogether, not because they couldn't obtain them, but to accelerate development of alternatives.
Huawei's Ascend chips became the centerpiece of this effort. Bernstein Research projects Huawei will claim 50% of China's AI chip market by 2026, with NVIDIA's share shrinking to 8%.
The performance gap remains real. Huawei's Ascend 910C achieves roughly 60-80% of NVIDIA H100's inference performance. But Chinese companies compensate through scale. Huawei's CloudMatrix 384 system links large numbers of chips to maximize collective performance. As Huawei's Xu Zhijun acknowledged: "Our single chip has a gap with NVIDIA, but through long-term investment in connection technology, our supernode can become the world's strongest."
The attitude shift extended to Jensen Huang himself. In July 2024, he suggested Huawei chips lagged NVIDIA by "10 years or more." A year later, he called Huawei "one of the most formidable technology companies in the world" and acknowledged replacement was "only a matter of time."
The TSMC Bottleneck
Ironically, both NVIDIA and Huawei depend on Taiwan's TSMC for manufacturing. NVIDIA approached TSMC to ramp up H200 production for Chinese orders. TSMC began volume production of 2-nanometer chips in late 2025.
But capacity constraints bite everyone. The primary bottleneck isn't chip fabrication but advanced packaging (CoWoS). NVIDIA needs what analysts called a "supply chain miracle" to meet Chinese demand while also serving hyperscalers elsewhere.
This creates a peculiar alignment: NVIDIA lobbies Washington for access to Chinese customers while China develops alternatives to reduce dependence on American technology, both companies reliant on Taiwanese manufacturing that neither fully controls.
The Contradictions Compound
The resulting landscape contains multiple paradoxes. The U.S. restricts chip exports for national security while collecting fees on those same exports. China bans foreign chips in state data centers while private companies scramble to order millions of American GPUs. Prosecutors charge smugglers with threatening national security one day; the president legalizes the same transactions the next.
Jensen Huang captures the tension. He visits Beijing and Washington in quick succession. He praises Huawei as a worthy competitor while lobbying for permission to compete against them. He watches China develop self-sufficiency partly because his company's chips became too difficult to obtain legally.
For NVIDIA, the tightrope walk continues. The company projects $8 billion in lost quarterly revenue from China restrictions. Chinese orders for 2026 far exceed current supply. Whether those chips ultimately ship depends on decisions in Washington, Beijing, Singapore, and Taipei that may change again before the containers are packed.
Links: CNBC: $160M GPU Smuggling (CNBC) | 观察者网: 华为挑战英伟达 (Guancha) | Taipei Times: NVIDIA-TSMC Orders (Taipei Times) | SCMP: H200 Shipments (SCMP) | IEEE Spectrum: China AI Chips (IEEE)